The answer probably lies in the clinical technologies that have been applied to the problem. Jha and colleagues [1] have dealt in depth with the two types of technology most likely to give useful data: prospective epidemiologic studies and randomized, controlled trials. The authors are correct in assuming that retrospective studies are too unreliable to give useful answers, and they do not review the many such studies available. Large numbers do not overcome serious deficiencies in experimental design. Jha and colleagues [1] found that the epidemiologic studies were usually more positive than the randomized, controlled trials, especially in the case of vitamin E, and they attribute this to the greater chance that unknown factors could be primarily responsible for the noted effects. It may also be that the epidemiologic studies were started first, when the control rates may have been higher, thus making it possible for the differences between the control and the experimental rates to be larger.

In fact, Jha and colleagues may have encountered a phenomenon that is likely to seal the fate of many therapeutic interventions. Large-scale randomized trials that are started late in the evolution of a therapy end up requiring still larger numbers of participants to prove efficacy because the control rates tend to decrease steadily with the passage of time. This statement is not referenced because, although it is common wisdom, I know of no meta-analyses that document it. The possible causes are many and include the more widespread application of generally beneficial treatments and the fact that practicing physicians in the later studies increasingly exclude from randomization the sicker patients, who might benefit from the newer therapies they have heard about. The controversy about intravenous magnesium in the treatment of acute myocardial infarction is a case in point. When early, small trials were combined in meta-analyses, the reduction in mortality was striking [2]. A larger trial, designed to be definitive, showed an appreciably smaller effect of magnesium and had a lower mortality rate in the control group [3]. The ISIS-4 trial [4], with more than 40000 patients, showed no effect of magnesium. This was in part because of delayed use of the drug [5] but was also, more impressively, a manifestation of the fact that the mortality rate in the control group was unusually low [6, 7].

Jha and colleagues [1] conclude that the epidemiologic evidence for the effectiveness of vitamin E is impressive but not sufficient to warrant action because of the common deficiencies of observational studies and because the dosages of vitamin E and the durations of therapy in the randomized, controlled trials were too small. The authors call for more sufficiently large randomized, controlled trials. I agree that these are necessary but provide a caveat that may not sit well with traditional frequentist biostatisticians: The new studies should not be done as if no previous data strongly suggest efficacy. They should be monitored closely by independent boards and should be stopped sooner than planned if the trends seem to be confirming that which has been previously found. Meta-analyses, which have been avoided for debatable reasons by Jha and colleagues [1], could be useful tools in determining when to stop [8].